Work Agent compaction follows the model window

What changed

Work Agent now aligns Claude Code's automatic-compaction threshold with the selected model's declared context window. Context windows from 100,000 through 1,000,000 tokens are used directly, while larger declared windows are capped at Claude Code's 1,000,000-token maximum.

Why this matters to the user

Long-context models no longer compact at Claude Code's generic 200,000-token default when Cherry Studio has more accurate model metadata. Editing that metadata also rebuilds the Work Agent connection before the next turn so the new threshold takes effect.

What the user should do

Nothing — this is automatic. For custom models, keep the model's context-window metadata accurate.

Notes for release manager

Claude Code rejects autoCompactWindow values below 100,000. Models below that limit, and models without context metadata, still use the SDK's 200,000-token default and need a separate host-side early-compaction solution.